Centralized Attention Modules for Hyperspectral Remote Sensing Image Classification

Hyperspectral image classification based upon the spectral-spatial information is a research hotspot in the field of remote sensing. The popular deep learning technique, attention mechanism, is generally used to search the influential bands and pixels for hyperspectral image classification. However, the spectral and spatial attention inferred by the existing attention modules may be weakly related to the center pixel, which is harmful to realize better classification. Inspired by the recent centralization ideology designed exclusively for hyperspectral image, the core components of three popular attention modules, including the dot-product part of self-attention module, the global semantic information of squeeze-and-excitation module, and the discovery of the relation between the global representation and crucial elements of convolutional block attention module, are improved. Therefore, the center pixel will be stimulated to make more contribution to the relevancy of attention. Experimental results on two classic data sets demonstrate the effectiveness of the centralized attention modules.